12 #include "lib/ieee754.h"
14 #include "lib/getrusage.h"
17 static struct fio_mutex *stat_mutex;
19 void update_rusage_stat(struct thread_data *td)
21 struct thread_stat *ts = &td->ts;
23 fio_getrusage(&td->ru_end);
24 ts->usr_time += mtime_since(&td->ru_start.ru_utime,
25 &td->ru_end.ru_utime);
26 ts->sys_time += mtime_since(&td->ru_start.ru_stime,
27 &td->ru_end.ru_stime);
28 ts->ctx += td->ru_end.ru_nvcsw + td->ru_end.ru_nivcsw
29 - (td->ru_start.ru_nvcsw + td->ru_start.ru_nivcsw);
30 ts->minf += td->ru_end.ru_minflt - td->ru_start.ru_minflt;
31 ts->majf += td->ru_end.ru_majflt - td->ru_start.ru_majflt;
33 memcpy(&td->ru_start, &td->ru_end, sizeof(td->ru_end));
37 * Given a latency, return the index of the corresponding bucket in
38 * the structure tracking percentiles.
40 * (1) find the group (and error bits) that the value (latency)
41 * belongs to by looking at its MSB. (2) find the bucket number in the
42 * group by looking at the index bits.
45 static unsigned int plat_val_to_idx(unsigned int val)
47 unsigned int msb, error_bits, base, offset, idx;
49 /* Find MSB starting from bit 0 */
53 msb = (sizeof(val)*8) - __builtin_clz(val) - 1;
56 * MSB <= (FIO_IO_U_PLAT_BITS-1), cannot be rounded off. Use
57 * all bits of the sample as index
59 if (msb <= FIO_IO_U_PLAT_BITS)
62 /* Compute the number of error bits to discard*/
63 error_bits = msb - FIO_IO_U_PLAT_BITS;
65 /* Compute the number of buckets before the group */
66 base = (error_bits + 1) << FIO_IO_U_PLAT_BITS;
69 * Discard the error bits and apply the mask to find the
70 * index for the buckets in the group
72 offset = (FIO_IO_U_PLAT_VAL - 1) & (val >> error_bits);
74 /* Make sure the index does not exceed (array size - 1) */
75 idx = (base + offset) < (FIO_IO_U_PLAT_NR - 1) ?
76 (base + offset) : (FIO_IO_U_PLAT_NR - 1);
82 * Convert the given index of the bucket array to the value
83 * represented by the bucket
85 static unsigned int plat_idx_to_val(unsigned int idx)
87 unsigned int error_bits, k, base;
89 assert(idx < FIO_IO_U_PLAT_NR);
91 /* MSB <= (FIO_IO_U_PLAT_BITS-1), cannot be rounded off. Use
92 * all bits of the sample as index */
93 if (idx < (FIO_IO_U_PLAT_VAL << 1))
96 /* Find the group and compute the minimum value of that group */
97 error_bits = (idx >> FIO_IO_U_PLAT_BITS) - 1;
98 base = 1 << (error_bits + FIO_IO_U_PLAT_BITS);
100 /* Find its bucket number of the group */
101 k = idx % FIO_IO_U_PLAT_VAL;
103 /* Return the mean of the range of the bucket */
104 return base + ((k + 0.5) * (1 << error_bits));
107 static int double_cmp(const void *a, const void *b)
109 const fio_fp64_t fa = *(const fio_fp64_t *) a;
110 const fio_fp64_t fb = *(const fio_fp64_t *) b;
115 else if (fa.u.f < fb.u.f)
121 unsigned int calc_clat_percentiles(unsigned int *io_u_plat, unsigned long nr,
122 fio_fp64_t *plist, unsigned int **output,
123 unsigned int *maxv, unsigned int *minv)
125 unsigned long sum = 0;
126 unsigned int len, i, j = 0;
127 unsigned int oval_len = 0;
128 unsigned int *ovals = NULL;
135 while (len < FIO_IO_U_LIST_MAX_LEN && plist[len].u.f != 0.0)
142 * Sort the percentile list. Note that it may already be sorted if
143 * we are using the default values, but since it's a short list this
144 * isn't a worry. Also note that this does not work for NaN values.
147 qsort((void *)plist, len, sizeof(plist[0]), double_cmp);
150 * Calculate bucket values, note down max and min values
153 for (i = 0; i < FIO_IO_U_PLAT_NR && !is_last; i++) {
155 while (sum >= (plist[j].u.f / 100.0 * nr)) {
156 assert(plist[j].u.f <= 100.0);
160 ovals = realloc(ovals, oval_len * sizeof(unsigned int));
163 ovals[j] = plat_idx_to_val(i);
164 if (ovals[j] < *minv)
166 if (ovals[j] > *maxv)
169 is_last = (j == len - 1);
182 * Find and display the p-th percentile of clat
184 static void show_clat_percentiles(unsigned int *io_u_plat, unsigned long nr,
185 fio_fp64_t *plist, unsigned int precision)
187 unsigned int len, j = 0, minv, maxv;
189 int is_last, per_line, scale_down;
192 len = calc_clat_percentiles(io_u_plat, nr, plist, &ovals, &maxv, &minv);
197 * We default to usecs, but if the value range is such that we
198 * should scale down to msecs, do that.
200 if (minv > 2000 && maxv > 99999) {
202 log_info(" clat percentiles (msec):\n |");
205 log_info(" clat percentiles (usec):\n |");
208 snprintf(fmt, sizeof(fmt), "%%1.%uf", precision);
209 per_line = (80 - 7) / (precision + 14);
211 for (j = 0; j < len; j++) {
212 char fbuf[16], *ptr = fbuf;
215 if (j != 0 && (j % per_line) == 0)
218 /* end of the list */
219 is_last = (j == len - 1);
221 if (plist[j].u.f < 10.0)
222 ptr += sprintf(fbuf, " ");
224 snprintf(ptr, sizeof(fbuf), fmt, plist[j].u.f);
227 ovals[j] = (ovals[j] + 999) / 1000;
229 log_info(" %sth=[%5u]%c", fbuf, ovals[j], is_last ? '\n' : ',');
234 if ((j % per_line) == per_line - 1) /* for formatting */
243 int calc_lat(struct io_stat *is, unsigned long *min, unsigned long *max,
244 double *mean, double *dev)
246 double n = (double) is->samples;
253 *mean = is->mean.u.f;
256 *dev = sqrt(is->S.u.f / (n - 1.0));
263 void show_group_stats(struct group_run_stats *rs)
265 char *p1, *p2, *p3, *p4;
266 const char *ddir_str[] = { " READ", " WRITE" , " TRIM"};
269 log_info("\nRun status group %d (all jobs):\n", rs->groupid);
271 for (i = 0; i < DDIR_RWDIR_CNT; i++) {
272 const int i2p = is_power_of_2(rs->kb_base);
277 p1 = num2str(rs->io_kb[i], 6, rs->kb_base, i2p, 8);
278 p2 = num2str(rs->agg[i], 6, rs->kb_base, i2p, rs->unit_base);
279 p3 = num2str(rs->min_bw[i], 6, rs->kb_base, i2p, rs->unit_base);
280 p4 = num2str(rs->max_bw[i], 6, rs->kb_base, i2p, rs->unit_base);
282 log_info("%s: io=%s, aggrb=%s/s, minb=%s/s, maxb=%s/s,"
283 " mint=%llumsec, maxt=%llumsec\n",
284 rs->unified_rw_rep ? " MIXED" : ddir_str[i],
286 (unsigned long long) rs->min_run[i],
287 (unsigned long long) rs->max_run[i]);
296 void stat_calc_dist(unsigned int *map, unsigned long total, double *io_u_dist)
301 * Do depth distribution calculations
303 for (i = 0; i < FIO_IO_U_MAP_NR; i++) {
305 io_u_dist[i] = (double) map[i] / (double) total;
306 io_u_dist[i] *= 100.0;
307 if (io_u_dist[i] < 0.1 && map[i])
314 static void stat_calc_lat(struct thread_stat *ts, double *dst,
315 unsigned int *src, int nr)
317 unsigned long total = ddir_rw_sum(ts->total_io_u);
321 * Do latency distribution calculations
323 for (i = 0; i < nr; i++) {
325 dst[i] = (double) src[i] / (double) total;
327 if (dst[i] < 0.01 && src[i])
334 void stat_calc_lat_u(struct thread_stat *ts, double *io_u_lat)
336 stat_calc_lat(ts, io_u_lat, ts->io_u_lat_u, FIO_IO_U_LAT_U_NR);
339 void stat_calc_lat_m(struct thread_stat *ts, double *io_u_lat)
341 stat_calc_lat(ts, io_u_lat, ts->io_u_lat_m, FIO_IO_U_LAT_M_NR);
344 static void display_lat(const char *name, unsigned long min, unsigned long max,
345 double mean, double dev)
347 const char *base = "(usec)";
350 if (!usec_to_msec(&min, &max, &mean, &dev))
353 minp = num2str(min, 6, 1, 0, 0);
354 maxp = num2str(max, 6, 1, 0, 0);
356 log_info(" %s %s: min=%s, max=%s, avg=%5.02f,"
357 " stdev=%5.02f\n", name, base, minp, maxp, mean, dev);
363 static void show_ddir_status(struct group_run_stats *rs, struct thread_stat *ts,
366 const char *ddir_str[] = { "read ", "write", "trim" };
367 unsigned long min, max, runt;
368 unsigned long long bw, iops;
370 char *io_p, *bw_p, *iops_p;
373 assert(ddir_rw(ddir));
375 if (!ts->runtime[ddir])
378 i2p = is_power_of_2(rs->kb_base);
379 runt = ts->runtime[ddir];
381 bw = (1000 * ts->io_bytes[ddir]) / runt;
382 io_p = num2str(ts->io_bytes[ddir], 6, 1, i2p, 8);
383 bw_p = num2str(bw, 6, 1, i2p, ts->unit_base);
385 iops = (1000 * (uint64_t)ts->total_io_u[ddir]) / runt;
386 iops_p = num2str(iops, 6, 1, 0, 0);
388 log_info(" %s: io=%s, bw=%s/s, iops=%s, runt=%6llumsec\n",
389 rs->unified_rw_rep ? "mixed" : ddir_str[ddir],
391 (unsigned long long) ts->runtime[ddir]);
397 if (calc_lat(&ts->slat_stat[ddir], &min, &max, &mean, &dev))
398 display_lat("slat", min, max, mean, dev);
399 if (calc_lat(&ts->clat_stat[ddir], &min, &max, &mean, &dev))
400 display_lat("clat", min, max, mean, dev);
401 if (calc_lat(&ts->lat_stat[ddir], &min, &max, &mean, &dev))
402 display_lat(" lat", min, max, mean, dev);
404 if (ts->clat_percentiles) {
405 show_clat_percentiles(ts->io_u_plat[ddir],
406 ts->clat_stat[ddir].samples,
408 ts->percentile_precision);
410 if (calc_lat(&ts->bw_stat[ddir], &min, &max, &mean, &dev)) {
411 double p_of_agg = 100.0, fkb_base = (double)rs->kb_base;
412 const char *bw_str = (rs->unit_base == 1 ? "Kbit" : "KB");
414 if (rs->unit_base == 1) {
422 p_of_agg = mean * 100 / (double) rs->agg[ddir];
423 if (p_of_agg > 100.0)
427 if (mean > fkb_base * fkb_base) {
432 bw_str = (rs->unit_base == 1 ? "Mbit" : "MB");
435 log_info(" bw (%-4s/s): min=%5lu, max=%5lu, per=%3.2f%%,"
436 " avg=%5.02f, stdev=%5.02f\n", bw_str, min, max,
437 p_of_agg, mean, dev);
441 static int show_lat(double *io_u_lat, int nr, const char **ranges,
444 int new_line = 1, i, line = 0, shown = 0;
446 for (i = 0; i < nr; i++) {
447 if (io_u_lat[i] <= 0.0)
453 log_info(" lat (%s) : ", msg);
459 log_info("%s%3.2f%%", ranges[i], io_u_lat[i]);
471 static void show_lat_u(double *io_u_lat_u)
473 const char *ranges[] = { "2=", "4=", "10=", "20=", "50=", "100=",
474 "250=", "500=", "750=", "1000=", };
476 show_lat(io_u_lat_u, FIO_IO_U_LAT_U_NR, ranges, "usec");
479 static void show_lat_m(double *io_u_lat_m)
481 const char *ranges[] = { "2=", "4=", "10=", "20=", "50=", "100=",
482 "250=", "500=", "750=", "1000=", "2000=",
485 show_lat(io_u_lat_m, FIO_IO_U_LAT_M_NR, ranges, "msec");
488 static void show_latencies(struct thread_stat *ts)
490 double io_u_lat_u[FIO_IO_U_LAT_U_NR];
491 double io_u_lat_m[FIO_IO_U_LAT_M_NR];
493 stat_calc_lat_u(ts, io_u_lat_u);
494 stat_calc_lat_m(ts, io_u_lat_m);
496 show_lat_u(io_u_lat_u);
497 show_lat_m(io_u_lat_m);
500 static void show_thread_status_normal(struct thread_stat *ts,
501 struct group_run_stats *rs)
503 double usr_cpu, sys_cpu;
504 unsigned long runtime;
505 double io_u_dist[FIO_IO_U_MAP_NR];
509 if (!(ts->io_bytes[DDIR_READ] + ts->io_bytes[DDIR_WRITE] +
510 ts->io_bytes[DDIR_TRIM]) && !(ts->total_io_u[DDIR_READ] +
511 ts->total_io_u[DDIR_WRITE] + ts->total_io_u[DDIR_TRIM]))
515 os_ctime_r((const time_t *) &time_p, time_buf, sizeof(time_buf));
518 log_info("%s: (groupid=%d, jobs=%d): err=%2d: pid=%d: %s",
519 ts->name, ts->groupid, ts->members,
520 ts->error, (int) ts->pid, time_buf);
522 log_info("%s: (groupid=%d, jobs=%d): err=%2d (%s): pid=%d: %s",
523 ts->name, ts->groupid, ts->members,
524 ts->error, ts->verror, (int) ts->pid,
528 if (strlen(ts->description))
529 log_info(" Description : [%s]\n", ts->description);
531 if (ts->io_bytes[DDIR_READ])
532 show_ddir_status(rs, ts, DDIR_READ);
533 if (ts->io_bytes[DDIR_WRITE])
534 show_ddir_status(rs, ts, DDIR_WRITE);
535 if (ts->io_bytes[DDIR_TRIM])
536 show_ddir_status(rs, ts, DDIR_TRIM);
540 runtime = ts->total_run_time;
542 double runt = (double) runtime;
544 usr_cpu = (double) ts->usr_time * 100 / runt;
545 sys_cpu = (double) ts->sys_time * 100 / runt;
551 log_info(" cpu : usr=%3.2f%%, sys=%3.2f%%, ctx=%llu,"
552 " majf=%llu, minf=%llu\n", usr_cpu, sys_cpu,
553 (unsigned long long) ts->ctx,
554 (unsigned long long) ts->majf,
555 (unsigned long long) ts->minf);
557 stat_calc_dist(ts->io_u_map, ddir_rw_sum(ts->total_io_u), io_u_dist);
558 log_info(" IO depths : 1=%3.1f%%, 2=%3.1f%%, 4=%3.1f%%, 8=%3.1f%%,"
559 " 16=%3.1f%%, 32=%3.1f%%, >=64=%3.1f%%\n", io_u_dist[0],
560 io_u_dist[1], io_u_dist[2],
561 io_u_dist[3], io_u_dist[4],
562 io_u_dist[5], io_u_dist[6]);
564 stat_calc_dist(ts->io_u_submit, ts->total_submit, io_u_dist);
565 log_info(" submit : 0=%3.1f%%, 4=%3.1f%%, 8=%3.1f%%, 16=%3.1f%%,"
566 " 32=%3.1f%%, 64=%3.1f%%, >=64=%3.1f%%\n", io_u_dist[0],
567 io_u_dist[1], io_u_dist[2],
568 io_u_dist[3], io_u_dist[4],
569 io_u_dist[5], io_u_dist[6]);
570 stat_calc_dist(ts->io_u_complete, ts->total_complete, io_u_dist);
571 log_info(" complete : 0=%3.1f%%, 4=%3.1f%%, 8=%3.1f%%, 16=%3.1f%%,"
572 " 32=%3.1f%%, 64=%3.1f%%, >=64=%3.1f%%\n", io_u_dist[0],
573 io_u_dist[1], io_u_dist[2],
574 io_u_dist[3], io_u_dist[4],
575 io_u_dist[5], io_u_dist[6]);
576 log_info(" issued : total=r=%llu/w=%llu/d=%llu,"
577 " short=r=%llu/w=%llu/d=%llu,"
578 " drop=r=%llu/w=%llu/d=%llu\n",
579 (unsigned long long) ts->total_io_u[0],
580 (unsigned long long) ts->total_io_u[1],
581 (unsigned long long) ts->total_io_u[2],
582 (unsigned long long) ts->short_io_u[0],
583 (unsigned long long) ts->short_io_u[1],
584 (unsigned long long) ts->short_io_u[2],
585 (unsigned long long) ts->drop_io_u[0],
586 (unsigned long long) ts->drop_io_u[1],
587 (unsigned long long) ts->drop_io_u[2]);
588 if (ts->continue_on_error) {
589 log_info(" errors : total=%llu, first_error=%d/<%s>\n",
590 (unsigned long long)ts->total_err_count,
592 strerror(ts->first_error));
594 if (ts->latency_depth) {
595 log_info(" latency : target=%llu, window=%llu, percentile=%.2f%%, depth=%u\n",
596 (unsigned long long)ts->latency_target,
597 (unsigned long long)ts->latency_window,
598 ts->latency_percentile.u.f,
603 static void show_ddir_status_terse(struct thread_stat *ts,
604 struct group_run_stats *rs, int ddir)
606 unsigned long min, max;
607 unsigned long long bw, iops;
608 unsigned int *ovals = NULL;
610 unsigned int len, minv, maxv;
613 assert(ddir_rw(ddir));
616 if (ts->runtime[ddir]) {
617 uint64_t runt = ts->runtime[ddir];
619 bw = ((1000 * ts->io_bytes[ddir]) / runt) / 1024;
620 iops = (1000 * (uint64_t) ts->total_io_u[ddir]) / runt;
623 log_info(";%llu;%llu;%llu;%llu",
624 (unsigned long long) ts->io_bytes[ddir] >> 10, bw, iops,
625 (unsigned long long) ts->runtime[ddir]);
627 if (calc_lat(&ts->slat_stat[ddir], &min, &max, &mean, &dev))
628 log_info(";%lu;%lu;%f;%f", min, max, mean, dev);
630 log_info(";%lu;%lu;%f;%f", 0UL, 0UL, 0.0, 0.0);
632 if (calc_lat(&ts->clat_stat[ddir], &min, &max, &mean, &dev))
633 log_info(";%lu;%lu;%f;%f", min, max, mean, dev);
635 log_info(";%lu;%lu;%f;%f", 0UL, 0UL, 0.0, 0.0);
637 if (ts->clat_percentiles) {
638 len = calc_clat_percentiles(ts->io_u_plat[ddir],
639 ts->clat_stat[ddir].samples,
640 ts->percentile_list, &ovals, &maxv,
645 for (i = 0; i < FIO_IO_U_LIST_MAX_LEN; i++) {
650 log_info(";%f%%=%u", ts->percentile_list[i].u.f, ovals[i]);
653 if (calc_lat(&ts->lat_stat[ddir], &min, &max, &mean, &dev))
654 log_info(";%lu;%lu;%f;%f", min, max, mean, dev);
656 log_info(";%lu;%lu;%f;%f", 0UL, 0UL, 0.0, 0.0);
661 if (calc_lat(&ts->bw_stat[ddir], &min, &max, &mean, &dev)) {
662 double p_of_agg = 100.0;
665 p_of_agg = mean * 100 / (double) rs->agg[ddir];
666 if (p_of_agg > 100.0)
670 log_info(";%lu;%lu;%f%%;%f;%f", min, max, p_of_agg, mean, dev);
672 log_info(";%lu;%lu;%f%%;%f;%f", 0UL, 0UL, 0.0, 0.0, 0.0);
675 static void add_ddir_status_json(struct thread_stat *ts,
676 struct group_run_stats *rs, int ddir, struct json_object *parent)
678 unsigned long min, max;
679 unsigned long long bw, iops;
680 unsigned int *ovals = NULL;
682 unsigned int len, minv, maxv;
684 const char *ddirname[] = {"read", "write", "trim"};
685 struct json_object *dir_object, *tmp_object, *percentile_object;
687 double p_of_agg = 100.0;
689 assert(ddir_rw(ddir));
691 if (ts->unified_rw_rep && ddir != DDIR_READ)
694 dir_object = json_create_object();
695 json_object_add_value_object(parent,
696 ts->unified_rw_rep ? "mixed" : ddirname[ddir], dir_object);
699 if (ts->runtime[ddir]) {
700 uint64_t runt = ts->runtime[ddir];
702 bw = ((1000 * ts->io_bytes[ddir]) / runt) / 1024;
703 iops = (1000 * (uint64_t) ts->total_io_u[ddir]) / runt;
706 json_object_add_value_int(dir_object, "io_bytes", ts->io_bytes[ddir] >> 10);
707 json_object_add_value_int(dir_object, "bw", bw);
708 json_object_add_value_int(dir_object, "iops", iops);
709 json_object_add_value_int(dir_object, "runtime", ts->runtime[ddir]);
711 if (!calc_lat(&ts->slat_stat[ddir], &min, &max, &mean, &dev)) {
715 tmp_object = json_create_object();
716 json_object_add_value_object(dir_object, "slat", tmp_object);
717 json_object_add_value_int(tmp_object, "min", min);
718 json_object_add_value_int(tmp_object, "max", max);
719 json_object_add_value_float(tmp_object, "mean", mean);
720 json_object_add_value_float(tmp_object, "stddev", dev);
722 if (!calc_lat(&ts->clat_stat[ddir], &min, &max, &mean, &dev)) {
726 tmp_object = json_create_object();
727 json_object_add_value_object(dir_object, "clat", tmp_object);
728 json_object_add_value_int(tmp_object, "min", min);
729 json_object_add_value_int(tmp_object, "max", max);
730 json_object_add_value_float(tmp_object, "mean", mean);
731 json_object_add_value_float(tmp_object, "stddev", dev);
733 if (ts->clat_percentiles) {
734 len = calc_clat_percentiles(ts->io_u_plat[ddir],
735 ts->clat_stat[ddir].samples,
736 ts->percentile_list, &ovals, &maxv,
741 percentile_object = json_create_object();
742 json_object_add_value_object(tmp_object, "percentile", percentile_object);
743 for (i = 0; i < FIO_IO_U_LIST_MAX_LEN; i++) {
745 json_object_add_value_int(percentile_object, "0.00", 0);
748 snprintf(buf, sizeof(buf), "%f", ts->percentile_list[i].u.f);
749 json_object_add_value_int(percentile_object, (const char *)buf, ovals[i]);
752 if (!calc_lat(&ts->lat_stat[ddir], &min, &max, &mean, &dev)) {
756 tmp_object = json_create_object();
757 json_object_add_value_object(dir_object, "lat", tmp_object);
758 json_object_add_value_int(tmp_object, "min", min);
759 json_object_add_value_int(tmp_object, "max", max);
760 json_object_add_value_float(tmp_object, "mean", mean);
761 json_object_add_value_float(tmp_object, "stddev", dev);
765 if (calc_lat(&ts->bw_stat[ddir], &min, &max, &mean, &dev)) {
767 p_of_agg = mean * 100 / (double) rs->agg[ddir];
768 if (p_of_agg > 100.0)
773 p_of_agg = mean = dev = 0.0;
775 json_object_add_value_int(dir_object, "bw_min", min);
776 json_object_add_value_int(dir_object, "bw_max", max);
777 json_object_add_value_float(dir_object, "bw_agg", p_of_agg);
778 json_object_add_value_float(dir_object, "bw_mean", mean);
779 json_object_add_value_float(dir_object, "bw_dev", dev);
782 static void show_thread_status_terse_v2(struct thread_stat *ts,
783 struct group_run_stats *rs)
785 double io_u_dist[FIO_IO_U_MAP_NR];
786 double io_u_lat_u[FIO_IO_U_LAT_U_NR];
787 double io_u_lat_m[FIO_IO_U_LAT_M_NR];
788 double usr_cpu, sys_cpu;
792 log_info("2;%s;%d;%d", ts->name, ts->groupid, ts->error);
793 /* Log Read Status */
794 show_ddir_status_terse(ts, rs, DDIR_READ);
795 /* Log Write Status */
796 show_ddir_status_terse(ts, rs, DDIR_WRITE);
797 /* Log Trim Status */
798 show_ddir_status_terse(ts, rs, DDIR_TRIM);
801 if (ts->total_run_time) {
802 double runt = (double) ts->total_run_time;
804 usr_cpu = (double) ts->usr_time * 100 / runt;
805 sys_cpu = (double) ts->sys_time * 100 / runt;
811 log_info(";%f%%;%f%%;%llu;%llu;%llu", usr_cpu, sys_cpu,
812 (unsigned long long) ts->ctx,
813 (unsigned long long) ts->majf,
814 (unsigned long long) ts->minf);
816 /* Calc % distribution of IO depths, usecond, msecond latency */
817 stat_calc_dist(ts->io_u_map, ddir_rw_sum(ts->total_io_u), io_u_dist);
818 stat_calc_lat_u(ts, io_u_lat_u);
819 stat_calc_lat_m(ts, io_u_lat_m);
821 /* Only show fixed 7 I/O depth levels*/
822 log_info(";%3.1f%%;%3.1f%%;%3.1f%%;%3.1f%%;%3.1f%%;%3.1f%%;%3.1f%%",
823 io_u_dist[0], io_u_dist[1], io_u_dist[2], io_u_dist[3],
824 io_u_dist[4], io_u_dist[5], io_u_dist[6]);
826 /* Microsecond latency */
827 for (i = 0; i < FIO_IO_U_LAT_U_NR; i++)
828 log_info(";%3.2f%%", io_u_lat_u[i]);
829 /* Millisecond latency */
830 for (i = 0; i < FIO_IO_U_LAT_M_NR; i++)
831 log_info(";%3.2f%%", io_u_lat_m[i]);
832 /* Additional output if continue_on_error set - default off*/
833 if (ts->continue_on_error)
834 log_info(";%llu;%d", (unsigned long long) ts->total_err_count, ts->first_error);
837 /* Additional output if description is set */
838 if (strlen(ts->description))
839 log_info(";%s", ts->description);
844 static void show_thread_status_terse_v3_v4(struct thread_stat *ts,
845 struct group_run_stats *rs, int ver)
847 double io_u_dist[FIO_IO_U_MAP_NR];
848 double io_u_lat_u[FIO_IO_U_LAT_U_NR];
849 double io_u_lat_m[FIO_IO_U_LAT_M_NR];
850 double usr_cpu, sys_cpu;
854 log_info("%d;%s;%s;%d;%d", ver, fio_version_string,
855 ts->name, ts->groupid, ts->error);
856 /* Log Read Status */
857 show_ddir_status_terse(ts, rs, DDIR_READ);
858 /* Log Write Status */
859 show_ddir_status_terse(ts, rs, DDIR_WRITE);
860 /* Log Trim Status */
862 show_ddir_status_terse(ts, rs, DDIR_TRIM);
865 if (ts->total_run_time) {
866 double runt = (double) ts->total_run_time;
868 usr_cpu = (double) ts->usr_time * 100 / runt;
869 sys_cpu = (double) ts->sys_time * 100 / runt;
875 log_info(";%f%%;%f%%;%llu;%llu;%llu", usr_cpu, sys_cpu,
876 (unsigned long long) ts->ctx,
877 (unsigned long long) ts->majf,
878 (unsigned long long) ts->minf);
880 /* Calc % distribution of IO depths, usecond, msecond latency */
881 stat_calc_dist(ts->io_u_map, ddir_rw_sum(ts->total_io_u), io_u_dist);
882 stat_calc_lat_u(ts, io_u_lat_u);
883 stat_calc_lat_m(ts, io_u_lat_m);
885 /* Only show fixed 7 I/O depth levels*/
886 log_info(";%3.1f%%;%3.1f%%;%3.1f%%;%3.1f%%;%3.1f%%;%3.1f%%;%3.1f%%",
887 io_u_dist[0], io_u_dist[1], io_u_dist[2], io_u_dist[3],
888 io_u_dist[4], io_u_dist[5], io_u_dist[6]);
890 /* Microsecond latency */
891 for (i = 0; i < FIO_IO_U_LAT_U_NR; i++)
892 log_info(";%3.2f%%", io_u_lat_u[i]);
893 /* Millisecond latency */
894 for (i = 0; i < FIO_IO_U_LAT_M_NR; i++)
895 log_info(";%3.2f%%", io_u_lat_m[i]);
897 /* disk util stats, if any */
899 show_disk_util(1, NULL);
901 /* Additional output if continue_on_error set - default off*/
902 if (ts->continue_on_error)
903 log_info(";%llu;%d", (unsigned long long) ts->total_err_count, ts->first_error);
905 /* Additional output if description is set */
906 if (strlen(ts->description))
907 log_info(";%s", ts->description);
912 static struct json_object *show_thread_status_json(struct thread_stat *ts,
913 struct group_run_stats *rs)
915 struct json_object *root, *tmp;
916 double io_u_dist[FIO_IO_U_MAP_NR];
917 double io_u_lat_u[FIO_IO_U_LAT_U_NR];
918 double io_u_lat_m[FIO_IO_U_LAT_M_NR];
919 double usr_cpu, sys_cpu;
922 root = json_create_object();
923 json_object_add_value_string(root, "jobname", ts->name);
924 json_object_add_value_int(root, "groupid", ts->groupid);
925 json_object_add_value_int(root, "error", ts->error);
927 add_ddir_status_json(ts, rs, DDIR_READ, root);
928 add_ddir_status_json(ts, rs, DDIR_WRITE, root);
929 add_ddir_status_json(ts, rs, DDIR_TRIM, root);
932 if (ts->total_run_time) {
933 double runt = (double) ts->total_run_time;
935 usr_cpu = (double) ts->usr_time * 100 / runt;
936 sys_cpu = (double) ts->sys_time * 100 / runt;
941 json_object_add_value_float(root, "usr_cpu", usr_cpu);
942 json_object_add_value_float(root, "sys_cpu", sys_cpu);
943 json_object_add_value_int(root, "ctx", ts->ctx);
944 json_object_add_value_int(root, "majf", ts->majf);
945 json_object_add_value_int(root, "minf", ts->minf);
948 /* Calc % distribution of IO depths, usecond, msecond latency */
949 stat_calc_dist(ts->io_u_map, ddir_rw_sum(ts->total_io_u), io_u_dist);
950 stat_calc_lat_u(ts, io_u_lat_u);
951 stat_calc_lat_m(ts, io_u_lat_m);
953 tmp = json_create_object();
954 json_object_add_value_object(root, "iodepth_level", tmp);
955 /* Only show fixed 7 I/O depth levels*/
956 for (i = 0; i < 7; i++) {
959 snprintf(name, 20, "%d", 1 << i);
961 snprintf(name, 20, ">=%d", 1 << i);
962 json_object_add_value_float(tmp, (const char *)name, io_u_dist[i]);
965 tmp = json_create_object();
966 json_object_add_value_object(root, "latency_us", tmp);
967 /* Microsecond latency */
968 for (i = 0; i < FIO_IO_U_LAT_U_NR; i++) {
969 const char *ranges[] = { "2", "4", "10", "20", "50", "100",
970 "250", "500", "750", "1000", };
971 json_object_add_value_float(tmp, ranges[i], io_u_lat_u[i]);
973 /* Millisecond latency */
974 tmp = json_create_object();
975 json_object_add_value_object(root, "latency_ms", tmp);
976 for (i = 0; i < FIO_IO_U_LAT_M_NR; i++) {
977 const char *ranges[] = { "2", "4", "10", "20", "50", "100",
978 "250", "500", "750", "1000", "2000",
980 json_object_add_value_float(tmp, ranges[i], io_u_lat_m[i]);
983 /* Additional output if continue_on_error set - default off*/
984 if (ts->continue_on_error) {
985 json_object_add_value_int(root, "total_err", ts->total_err_count);
986 json_object_add_value_int(root, "first_error", ts->first_error);
989 if (ts->latency_depth) {
990 json_object_add_value_int(root, "latency_depth", ts->latency_depth);
991 json_object_add_value_int(root, "latency_target", ts->latency_target);
992 json_object_add_value_float(root, "latency_percentile", ts->latency_percentile.u.f);
993 json_object_add_value_int(root, "latency_window", ts->latency_window);
996 /* Additional output if description is set */
997 if (strlen(ts->description))
998 json_object_add_value_string(root, "desc", ts->description);
1003 static void show_thread_status_terse(struct thread_stat *ts,
1004 struct group_run_stats *rs)
1006 if (terse_version == 2)
1007 show_thread_status_terse_v2(ts, rs);
1008 else if (terse_version == 3 || terse_version == 4)
1009 show_thread_status_terse_v3_v4(ts, rs, terse_version);
1011 log_err("fio: bad terse version!? %d\n", terse_version);
1014 struct json_object *show_thread_status(struct thread_stat *ts,
1015 struct group_run_stats *rs)
1017 if (output_format == FIO_OUTPUT_TERSE)
1018 show_thread_status_terse(ts, rs);
1019 else if (output_format == FIO_OUTPUT_JSON)
1020 return show_thread_status_json(ts, rs);
1022 show_thread_status_normal(ts, rs);
1026 static void sum_stat(struct io_stat *dst, struct io_stat *src, int nr)
1030 if (src->samples == 0)
1033 dst->min_val = min(dst->min_val, src->min_val);
1034 dst->max_val = max(dst->max_val, src->max_val);
1037 * Compute new mean and S after the merge
1038 * <http://en.wikipedia.org/wiki/Algorithms_for_calculating_variance
1039 * #Parallel_algorithm>
1042 mean = src->mean.u.f;
1045 double delta = src->mean.u.f - dst->mean.u.f;
1047 mean = ((src->mean.u.f * src->samples) +
1048 (dst->mean.u.f * dst->samples)) /
1049 (dst->samples + src->samples);
1051 S = src->S.u.f + dst->S.u.f + pow(delta, 2.0) *
1052 (dst->samples * src->samples) /
1053 (dst->samples + src->samples);
1056 dst->samples += src->samples;
1057 dst->mean.u.f = mean;
1061 void sum_group_stats(struct group_run_stats *dst, struct group_run_stats *src)
1065 for (i = 0; i < DDIR_RWDIR_CNT; i++) {
1066 if (dst->max_run[i] < src->max_run[i])
1067 dst->max_run[i] = src->max_run[i];
1068 if (dst->min_run[i] && dst->min_run[i] > src->min_run[i])
1069 dst->min_run[i] = src->min_run[i];
1070 if (dst->max_bw[i] < src->max_bw[i])
1071 dst->max_bw[i] = src->max_bw[i];
1072 if (dst->min_bw[i] && dst->min_bw[i] > src->min_bw[i])
1073 dst->min_bw[i] = src->min_bw[i];
1075 dst->io_kb[i] += src->io_kb[i];
1076 dst->agg[i] += src->agg[i];
1081 void sum_thread_stats(struct thread_stat *dst, struct thread_stat *src, int nr)
1085 for (l = 0; l < DDIR_RWDIR_CNT; l++) {
1086 if (!dst->unified_rw_rep) {
1087 sum_stat(&dst->clat_stat[l], &src->clat_stat[l], nr);
1088 sum_stat(&dst->slat_stat[l], &src->slat_stat[l], nr);
1089 sum_stat(&dst->lat_stat[l], &src->lat_stat[l], nr);
1090 sum_stat(&dst->bw_stat[l], &src->bw_stat[l], nr);
1092 dst->io_bytes[l] += src->io_bytes[l];
1094 if (dst->runtime[l] < src->runtime[l])
1095 dst->runtime[l] = src->runtime[l];
1097 sum_stat(&dst->clat_stat[0], &src->clat_stat[l], nr);
1098 sum_stat(&dst->slat_stat[0], &src->slat_stat[l], nr);
1099 sum_stat(&dst->lat_stat[0], &src->lat_stat[l], nr);
1100 sum_stat(&dst->bw_stat[0], &src->bw_stat[l], nr);
1102 dst->io_bytes[0] += src->io_bytes[l];
1104 if (dst->runtime[0] < src->runtime[l])
1105 dst->runtime[0] = src->runtime[l];
1109 dst->usr_time += src->usr_time;
1110 dst->sys_time += src->sys_time;
1111 dst->ctx += src->ctx;
1112 dst->majf += src->majf;
1113 dst->minf += src->minf;
1115 for (k = 0; k < FIO_IO_U_MAP_NR; k++)
1116 dst->io_u_map[k] += src->io_u_map[k];
1117 for (k = 0; k < FIO_IO_U_MAP_NR; k++)
1118 dst->io_u_submit[k] += src->io_u_submit[k];
1119 for (k = 0; k < FIO_IO_U_MAP_NR; k++)
1120 dst->io_u_complete[k] += src->io_u_complete[k];
1121 for (k = 0; k < FIO_IO_U_LAT_U_NR; k++)
1122 dst->io_u_lat_u[k] += src->io_u_lat_u[k];
1123 for (k = 0; k < FIO_IO_U_LAT_M_NR; k++)
1124 dst->io_u_lat_m[k] += src->io_u_lat_m[k];
1126 for (k = 0; k < DDIR_RWDIR_CNT; k++) {
1127 if (!dst->unified_rw_rep) {
1128 dst->total_io_u[k] += src->total_io_u[k];
1129 dst->short_io_u[k] += src->short_io_u[k];
1130 dst->drop_io_u[k] += src->drop_io_u[k];
1132 dst->total_io_u[0] += src->total_io_u[k];
1133 dst->short_io_u[0] += src->short_io_u[k];
1134 dst->drop_io_u[0] += src->drop_io_u[k];
1138 for (k = 0; k < DDIR_RWDIR_CNT; k++) {
1141 for (m = 0; m < FIO_IO_U_PLAT_NR; m++) {
1142 if (!dst->unified_rw_rep)
1143 dst->io_u_plat[k][m] += src->io_u_plat[k][m];
1145 dst->io_u_plat[0][m] += src->io_u_plat[k][m];
1149 dst->total_run_time += src->total_run_time;
1150 dst->total_submit += src->total_submit;
1151 dst->total_complete += src->total_complete;
1154 void init_group_run_stat(struct group_run_stats *gs)
1157 memset(gs, 0, sizeof(*gs));
1159 for (i = 0; i < DDIR_RWDIR_CNT; i++)
1160 gs->min_bw[i] = gs->min_run[i] = ~0UL;
1163 void init_thread_stat(struct thread_stat *ts)
1167 memset(ts, 0, sizeof(*ts));
1169 for (j = 0; j < DDIR_RWDIR_CNT; j++) {
1170 ts->lat_stat[j].min_val = -1UL;
1171 ts->clat_stat[j].min_val = -1UL;
1172 ts->slat_stat[j].min_val = -1UL;
1173 ts->bw_stat[j].min_val = -1UL;
1178 void __show_run_stats(void)
1180 struct group_run_stats *runstats, *rs;
1181 struct thread_data *td;
1182 struct thread_stat *threadstats, *ts;
1183 int i, j, nr_ts, last_ts, idx;
1184 int kb_base_warned = 0;
1185 int unit_base_warned = 0;
1186 struct json_object *root = NULL;
1187 struct json_array *array = NULL;
1189 runstats = malloc(sizeof(struct group_run_stats) * (groupid + 1));
1191 for (i = 0; i < groupid + 1; i++)
1192 init_group_run_stat(&runstats[i]);
1195 * find out how many threads stats we need. if group reporting isn't
1196 * enabled, it's one-per-td.
1200 for_each_td(td, i) {
1201 if (!td->o.group_reporting) {
1205 if (last_ts == td->groupid)
1208 last_ts = td->groupid;
1212 threadstats = malloc(nr_ts * sizeof(struct thread_stat));
1214 for (i = 0; i < nr_ts; i++)
1215 init_thread_stat(&threadstats[i]);
1220 for_each_td(td, i) {
1221 if (idx && (!td->o.group_reporting ||
1222 (td->o.group_reporting && last_ts != td->groupid))) {
1227 last_ts = td->groupid;
1229 ts = &threadstats[j];
1231 ts->clat_percentiles = td->o.clat_percentiles;
1232 ts->percentile_precision = td->o.percentile_precision;
1233 memcpy(ts->percentile_list, td->o.percentile_list, sizeof(td->o.percentile_list));
1238 if (ts->groupid == -1) {
1240 * These are per-group shared already
1242 strncpy(ts->name, td->o.name, FIO_JOBNAME_SIZE - 1);
1243 if (td->o.description)
1244 strncpy(ts->description, td->o.description,
1245 FIO_JOBDESC_SIZE - 1);
1247 memset(ts->description, 0, FIO_JOBDESC_SIZE);
1250 * If multiple entries in this group, this is
1253 ts->thread_number = td->thread_number;
1254 ts->groupid = td->groupid;
1257 * first pid in group, not very useful...
1261 ts->kb_base = td->o.kb_base;
1262 ts->unit_base = td->o.unit_base;
1263 ts->unified_rw_rep = td->o.unified_rw_rep;
1264 } else if (ts->kb_base != td->o.kb_base && !kb_base_warned) {
1265 log_info("fio: kb_base differs for jobs in group, using"
1266 " %u as the base\n", ts->kb_base);
1268 } else if (ts->unit_base != td->o.unit_base && !unit_base_warned) {
1269 log_info("fio: unit_base differs for jobs in group, using"
1270 " %u as the base\n", ts->unit_base);
1271 unit_base_warned = 1;
1274 ts->continue_on_error = td->o.continue_on_error;
1275 ts->total_err_count += td->total_err_count;
1276 ts->first_error = td->first_error;
1278 if (!td->error && td->o.continue_on_error &&
1280 ts->error = td->first_error;
1281 ts->verror[sizeof(ts->verror) - 1] = '\0';
1282 strncpy(ts->verror, td->verror, sizeof(ts->verror) - 1);
1283 } else if (td->error) {
1284 ts->error = td->error;
1285 ts->verror[sizeof(ts->verror) - 1] = '\0';
1286 strncpy(ts->verror, td->verror, sizeof(ts->verror) - 1);
1290 ts->latency_depth = td->latency_qd;
1291 ts->latency_target = td->o.latency_target;
1292 ts->latency_percentile = td->o.latency_percentile;
1293 ts->latency_window = td->o.latency_window;
1295 sum_thread_stats(ts, &td->ts, idx);
1298 for (i = 0; i < nr_ts; i++) {
1299 unsigned long long bw;
1301 ts = &threadstats[i];
1302 rs = &runstats[ts->groupid];
1303 rs->kb_base = ts->kb_base;
1304 rs->unit_base = ts->unit_base;
1305 rs->unified_rw_rep += ts->unified_rw_rep;
1307 for (j = 0; j < DDIR_RWDIR_CNT; j++) {
1308 if (!ts->runtime[j])
1310 if (ts->runtime[j] < rs->min_run[j] || !rs->min_run[j])
1311 rs->min_run[j] = ts->runtime[j];
1312 if (ts->runtime[j] > rs->max_run[j])
1313 rs->max_run[j] = ts->runtime[j];
1316 if (ts->runtime[j]) {
1317 unsigned long runt = ts->runtime[j];
1318 unsigned long long kb;
1320 kb = ts->io_bytes[j] / rs->kb_base;
1321 bw = kb * 1000 / runt;
1323 if (bw < rs->min_bw[j])
1325 if (bw > rs->max_bw[j])
1328 rs->io_kb[j] += ts->io_bytes[j] / rs->kb_base;
1332 for (i = 0; i < groupid + 1; i++) {
1337 for (ddir = 0; ddir < DDIR_RWDIR_CNT; ddir++) {
1338 if (rs->max_run[ddir])
1339 rs->agg[ddir] = (rs->io_kb[ddir] * 1000) /
1345 * don't overwrite last signal output
1347 if (output_format == FIO_OUTPUT_NORMAL)
1349 else if (output_format == FIO_OUTPUT_JSON) {
1350 root = json_create_object();
1351 json_object_add_value_string(root, "fio version", fio_version_string);
1352 array = json_create_array();
1353 json_object_add_value_array(root, "jobs", array);
1356 for (i = 0; i < nr_ts; i++) {
1357 ts = &threadstats[i];
1358 rs = &runstats[ts->groupid];
1361 fio_server_send_ts(ts, rs);
1362 else if (output_format == FIO_OUTPUT_TERSE)
1363 show_thread_status_terse(ts, rs);
1364 else if (output_format == FIO_OUTPUT_JSON) {
1365 struct json_object *tmp = show_thread_status_json(ts, rs);
1366 json_array_add_value_object(array, tmp);
1368 show_thread_status_normal(ts, rs);
1370 if (output_format == FIO_OUTPUT_JSON) {
1371 /* disk util stats, if any */
1372 show_disk_util(1, root);
1374 show_idle_prof_stats(FIO_OUTPUT_JSON, root);
1376 json_print_object(root);
1378 json_free_object(root);
1381 for (i = 0; i < groupid + 1; i++) {
1386 fio_server_send_gs(rs);
1387 else if (output_format == FIO_OUTPUT_NORMAL)
1388 show_group_stats(rs);
1392 fio_server_send_du();
1393 else if (output_format == FIO_OUTPUT_NORMAL) {
1394 show_disk_util(0, NULL);
1395 show_idle_prof_stats(FIO_OUTPUT_NORMAL, NULL);
1398 if ( !(output_format == FIO_OUTPUT_TERSE) && append_terse_output) {
1399 log_info("\nAdditional Terse Output:\n");
1401 for (i = 0; i < nr_ts; i++) {
1402 ts = &threadstats[i];
1403 rs = &runstats[ts->groupid];
1404 show_thread_status_terse(ts, rs);
1413 void show_run_stats(void)
1415 fio_mutex_down(stat_mutex);
1417 fio_mutex_up(stat_mutex);
1420 static void *__show_running_run_stats(void *arg)
1422 struct thread_data *td;
1423 unsigned long long *rt;
1427 fio_mutex_down(stat_mutex);
1429 rt = malloc(thread_number * sizeof(unsigned long long));
1430 fio_gettime(&tv, NULL);
1432 for_each_td(td, i) {
1433 rt[i] = mtime_since(&td->start, &tv);
1434 if (td_read(td) && td->io_bytes[DDIR_READ])
1435 td->ts.runtime[DDIR_READ] += rt[i];
1436 if (td_write(td) && td->io_bytes[DDIR_WRITE])
1437 td->ts.runtime[DDIR_WRITE] += rt[i];
1438 if (td_trim(td) && td->io_bytes[DDIR_TRIM])
1439 td->ts.runtime[DDIR_TRIM] += rt[i];
1441 td->update_rusage = 1;
1442 td->ts.io_bytes[DDIR_READ] = td->io_bytes[DDIR_READ];
1443 td->ts.io_bytes[DDIR_WRITE] = td->io_bytes[DDIR_WRITE];
1444 td->ts.io_bytes[DDIR_TRIM] = td->io_bytes[DDIR_TRIM];
1445 td->ts.total_run_time = mtime_since(&td->epoch, &tv);
1448 for_each_td(td, i) {
1449 if (td->rusage_sem) {
1450 td->update_rusage = 1;
1451 fio_mutex_down(td->rusage_sem);
1453 td->update_rusage = 0;
1458 for_each_td(td, i) {
1459 if (td_read(td) && td->io_bytes[DDIR_READ])
1460 td->ts.runtime[DDIR_READ] -= rt[i];
1461 if (td_write(td) && td->io_bytes[DDIR_WRITE])
1462 td->ts.runtime[DDIR_WRITE] -= rt[i];
1463 if (td_trim(td) && td->io_bytes[DDIR_TRIM])
1464 td->ts.runtime[DDIR_TRIM] -= rt[i];
1468 fio_mutex_up(stat_mutex);
1474 * Called from signal handler. It _should_ be safe to just run this inline
1475 * in the sig handler, but we should be disturbing the system less by just
1476 * creating a thread to do it.
1478 void show_running_run_stats(void)
1482 thread = calloc(1, sizeof(*thread));
1486 if (!pthread_create(thread, NULL, __show_running_run_stats, thread)) {
1489 err = pthread_detach(*thread);
1491 log_err("fio: DU thread detach failed: %s\n", strerror(err));
1499 static int status_interval_init;
1500 static struct timeval status_time;
1501 static int status_file_disabled;
1503 #define FIO_STATUS_FILE "fio-dump-status"
1505 static int check_status_file(void)
1508 const char *temp_dir;
1509 char fio_status_file_path[PATH_MAX];
1511 if (status_file_disabled)
1514 temp_dir = getenv("TMPDIR");
1515 if (temp_dir == NULL) {
1516 temp_dir = getenv("TEMP");
1517 if (temp_dir && strlen(temp_dir) >= PATH_MAX)
1520 if (temp_dir == NULL)
1523 snprintf(fio_status_file_path, sizeof(fio_status_file_path), "%s/%s", temp_dir, FIO_STATUS_FILE);
1525 if (stat(fio_status_file_path, &sb))
1528 if (unlink(fio_status_file_path) < 0) {
1529 log_err("fio: failed to unlink %s: %s\n", fio_status_file_path,
1531 log_err("fio: disabling status file updates\n");
1532 status_file_disabled = 1;
1538 void check_for_running_stats(void)
1540 if (status_interval) {
1541 if (!status_interval_init) {
1542 fio_gettime(&status_time, NULL);
1543 status_interval_init = 1;
1544 } else if (mtime_since_now(&status_time) >= status_interval) {
1545 show_running_run_stats();
1546 fio_gettime(&status_time, NULL);
1550 if (check_status_file()) {
1551 show_running_run_stats();
1556 static inline void add_stat_sample(struct io_stat *is, unsigned long data)
1561 if (data > is->max_val)
1563 if (data < is->min_val)
1566 delta = val - is->mean.u.f;
1568 is->mean.u.f += delta / (is->samples + 1.0);
1569 is->S.u.f += delta * (val - is->mean.u.f);
1575 static void __add_log_sample(struct io_log *iolog, unsigned long val,
1576 enum fio_ddir ddir, unsigned int bs,
1577 unsigned long t, uint64_t offset)
1579 uint64_t nr_samples = iolog->nr_samples;
1580 struct io_sample *s;
1582 if (iolog->disabled)
1585 if (!iolog->nr_samples)
1586 iolog->avg_last = t;
1588 if (iolog->nr_samples == iolog->max_samples) {
1592 new_size = 2 * iolog->max_samples * log_entry_sz(iolog);
1594 if (iolog->log_gz && (new_size > iolog->log_gz)) {
1595 if (iolog_flush(iolog, 0)) {
1596 log_err("fio: failed flushing iolog! Will stop logging.\n");
1597 iolog->disabled = 1;
1600 nr_samples = iolog->nr_samples;
1602 new_log = realloc(iolog->log, new_size);
1604 log_err("fio: failed extending iolog! Will stop logging.\n");
1605 iolog->disabled = 1;
1608 iolog->log = new_log;
1609 iolog->max_samples <<= 1;
1613 s = get_sample(iolog, nr_samples);
1617 io_sample_set_ddir(iolog, s, ddir);
1620 if (iolog->log_offset) {
1621 struct io_sample_offset *so = (void *) s;
1623 so->offset = offset;
1626 iolog->nr_samples++;
1629 static inline void reset_io_stat(struct io_stat *ios)
1631 ios->max_val = ios->min_val = ios->samples = 0;
1632 ios->mean.u.f = ios->S.u.f = 0;
1635 void reset_io_stats(struct thread_data *td)
1637 struct thread_stat *ts = &td->ts;
1640 for (i = 0; i < DDIR_RWDIR_CNT; i++) {
1641 reset_io_stat(&ts->clat_stat[i]);
1642 reset_io_stat(&ts->slat_stat[i]);
1643 reset_io_stat(&ts->lat_stat[i]);
1644 reset_io_stat(&ts->bw_stat[i]);
1645 reset_io_stat(&ts->iops_stat[i]);
1647 ts->io_bytes[i] = 0;
1650 for (j = 0; j < FIO_IO_U_PLAT_NR; j++)
1651 ts->io_u_plat[i][j] = 0;
1654 for (i = 0; i < FIO_IO_U_MAP_NR; i++) {
1655 ts->io_u_map[i] = 0;
1656 ts->io_u_submit[i] = 0;
1657 ts->io_u_complete[i] = 0;
1658 ts->io_u_lat_u[i] = 0;
1659 ts->io_u_lat_m[i] = 0;
1660 ts->total_submit = 0;
1661 ts->total_complete = 0;
1664 for (i = 0; i < 3; i++) {
1665 ts->total_io_u[i] = 0;
1666 ts->short_io_u[i] = 0;
1667 ts->drop_io_u[i] = 0;
1671 static void _add_stat_to_log(struct io_log *iolog, unsigned long elapsed)
1674 * Note an entry in the log. Use the mean from the logged samples,
1675 * making sure to properly round up. Only write a log entry if we
1676 * had actual samples done.
1678 if (iolog->avg_window[DDIR_READ].samples) {
1681 mr = iolog->avg_window[DDIR_READ].mean.u.f + 0.50;
1682 __add_log_sample(iolog, mr, DDIR_READ, 0, elapsed, 0);
1684 if (iolog->avg_window[DDIR_WRITE].samples) {
1687 mw = iolog->avg_window[DDIR_WRITE].mean.u.f + 0.50;
1688 __add_log_sample(iolog, mw, DDIR_WRITE, 0, elapsed, 0);
1690 if (iolog->avg_window[DDIR_TRIM].samples) {
1693 mw = iolog->avg_window[DDIR_TRIM].mean.u.f + 0.50;
1694 __add_log_sample(iolog, mw, DDIR_TRIM, 0, elapsed, 0);
1697 reset_io_stat(&iolog->avg_window[DDIR_READ]);
1698 reset_io_stat(&iolog->avg_window[DDIR_WRITE]);
1699 reset_io_stat(&iolog->avg_window[DDIR_TRIM]);
1702 static void add_log_sample(struct thread_data *td, struct io_log *iolog,
1703 unsigned long val, enum fio_ddir ddir,
1704 unsigned int bs, uint64_t offset)
1706 unsigned long elapsed, this_window;
1711 elapsed = mtime_since_now(&td->epoch);
1714 * If no time averaging, just add the log sample.
1716 if (!iolog->avg_msec) {
1717 __add_log_sample(iolog, val, ddir, bs, elapsed, offset);
1722 * Add the sample. If the time period has passed, then
1723 * add that entry to the log and clear.
1725 add_stat_sample(&iolog->avg_window[ddir], val);
1728 * If period hasn't passed, adding the above sample is all we
1731 this_window = elapsed - iolog->avg_last;
1732 if (this_window < iolog->avg_msec)
1735 _add_stat_to_log(iolog, elapsed);
1737 iolog->avg_last = elapsed;
1740 void finalize_logs(struct thread_data *td)
1742 unsigned long elapsed;
1744 elapsed = mtime_since_now(&td->epoch);
1747 _add_stat_to_log(td->clat_log, elapsed);
1749 _add_stat_to_log(td->slat_log, elapsed);
1751 _add_stat_to_log(td->lat_log, elapsed);
1753 _add_stat_to_log(td->bw_log, elapsed);
1755 _add_stat_to_log(td->iops_log, elapsed);
1758 void add_agg_sample(unsigned long val, enum fio_ddir ddir, unsigned int bs)
1760 struct io_log *iolog;
1765 iolog = agg_io_log[ddir];
1766 __add_log_sample(iolog, val, ddir, bs, mtime_since_genesis(), 0);
1769 static void add_clat_percentile_sample(struct thread_stat *ts,
1770 unsigned long usec, enum fio_ddir ddir)
1772 unsigned int idx = plat_val_to_idx(usec);
1773 assert(idx < FIO_IO_U_PLAT_NR);
1775 ts->io_u_plat[ddir][idx]++;
1778 void add_clat_sample(struct thread_data *td, enum fio_ddir ddir,
1779 unsigned long usec, unsigned int bs, uint64_t offset)
1781 struct thread_stat *ts = &td->ts;
1786 add_stat_sample(&ts->clat_stat[ddir], usec);
1789 add_log_sample(td, td->clat_log, usec, ddir, bs, offset);
1791 if (ts->clat_percentiles)
1792 add_clat_percentile_sample(ts, usec, ddir);
1795 void add_slat_sample(struct thread_data *td, enum fio_ddir ddir,
1796 unsigned long usec, unsigned int bs, uint64_t offset)
1798 struct thread_stat *ts = &td->ts;
1803 add_stat_sample(&ts->slat_stat[ddir], usec);
1806 add_log_sample(td, td->slat_log, usec, ddir, bs, offset);
1809 void add_lat_sample(struct thread_data *td, enum fio_ddir ddir,
1810 unsigned long usec, unsigned int bs, uint64_t offset)
1812 struct thread_stat *ts = &td->ts;
1817 add_stat_sample(&ts->lat_stat[ddir], usec);
1820 add_log_sample(td, td->lat_log, usec, ddir, bs, offset);
1823 void add_bw_sample(struct thread_data *td, enum fio_ddir ddir, unsigned int bs,
1826 struct thread_stat *ts = &td->ts;
1827 unsigned long spent, rate;
1832 spent = mtime_since(&td->bw_sample_time, t);
1833 if (spent < td->o.bw_avg_time)
1837 * Compute both read and write rates for the interval.
1839 for (ddir = DDIR_READ; ddir < DDIR_RWDIR_CNT; ddir++) {
1842 delta = td->this_io_bytes[ddir] - td->stat_io_bytes[ddir];
1844 continue; /* No entries for interval */
1847 rate = delta * 1000 / spent / 1024;
1851 add_stat_sample(&ts->bw_stat[ddir], rate);
1854 add_log_sample(td, td->bw_log, rate, ddir, bs, 0);
1856 td->stat_io_bytes[ddir] = td->this_io_bytes[ddir];
1859 fio_gettime(&td->bw_sample_time, NULL);
1862 void add_iops_sample(struct thread_data *td, enum fio_ddir ddir, unsigned int bs,
1865 struct thread_stat *ts = &td->ts;
1866 unsigned long spent, iops;
1871 spent = mtime_since(&td->iops_sample_time, t);
1872 if (spent < td->o.iops_avg_time)
1876 * Compute both read and write rates for the interval.
1878 for (ddir = DDIR_READ; ddir < DDIR_RWDIR_CNT; ddir++) {
1881 delta = td->this_io_blocks[ddir] - td->stat_io_blocks[ddir];
1883 continue; /* No entries for interval */
1886 iops = (delta * 1000) / spent;
1890 add_stat_sample(&ts->iops_stat[ddir], iops);
1893 add_log_sample(td, td->iops_log, iops, ddir, bs, 0);
1895 td->stat_io_blocks[ddir] = td->this_io_blocks[ddir];
1898 fio_gettime(&td->iops_sample_time, NULL);
1901 void stat_init(void)
1903 stat_mutex = fio_mutex_init(FIO_MUTEX_UNLOCKED);
1906 void stat_exit(void)
1909 * When we have the mutex, we know out-of-band access to it
1912 fio_mutex_down(stat_mutex);
1913 fio_mutex_remove(stat_mutex);